Bearing fault diagnosis method based on improved Siamese neural network with small sample

• DOI: 10.1186/s13677-022-00350-1
• Publication Date: 2022-11-19T01:02:48Z
• Authors: Xiaoping Zhao; Mengyao Ma; Fan Shao
• Keywords: Similarity (geometry); Sample (material); Feature (linguistics)

Abstract Fault diagnosis of rolling bearings is very important for monitoring the health rotating machinery. However, in actual industrial production, owing to constraints conditions and costs, only a small number bearing fault samples can be obtained, which leads an unsatisfactory effect traditional models based on data-driven methods. Therefore, this study proposes small-sample method improved Siamese neural network (ISNN). This adds classification branch standard replaces common Euclidean distance measurement with measurement. The model includes three networks: feature extraction network, relationship network. First, were input into same pairs, long short-term memory (LSTM) convolutional (CNN) used map signal data low-dimensional space. Then, extracted sample features measured similarity by network; at time, complete recognition. When training was particularly (training set A, 10 samples), accuracy 1D CNN, Prototype net 49.8%, 60.2% 58.6% respectively, while proposed ISNN 84.1%. For 100-sample case D, CNN 93.4%, still higher than that prototype Siam reached 98.1%.The experimental results show achieved better generalization samples.